Evaluating Visual Mathematics in Multimodal LLMs: A Multilingual Benchmark Based on the Kangaroo Tests

TL;DR

This paper evaluates the mathematical reasoning capabilities of multimodal large language models across multiple languages and mathematical domains, revealing moderate performance and highlighting areas for improvement in diagram understanding and reasoning.

Contribution

It introduces a multilingual benchmark for visual mathematics in MLLMs and provides a comprehensive analysis of several models' reasoning abilities and limitations.

Findings

Gemini 2.0 Flash achieves highest accuracy on image tasks.

Models perform better on easier questions but struggle with advanced reasoning.

Significant variation exists across languages and difficulty levels.

Abstract

Multimodal Large Language Models (MLLMs) promise advanced vision language capabilities, yet their effectiveness in visually presented mathematics remains underexplored. This paper analyzes the development and evaluation of MLLMs for mathematical problem solving, focusing on diagrams, multilingual text, and symbolic notation. We then assess several models, including GPT 4o, Pixtral, Qwen VL, Llama 3.2 Vision variants, and Gemini 2.0 Flash in a multilingual Kangaroo style benchmark spanning English, French, Spanish, and Catalan. Our experiments reveal four key findings. First, overall precision remains moderate across geometry, visual algebra, logic, patterns, and combinatorics: no single model excels in every topic. Second, while most models see improved accuracy with questions that do not have images, the gain is often limited; performance for some remains nearly unchanged without visual input, indicating underutilization of diagrammatic information. Third, substantial variation exists across languages and difficulty levels: models frequently handle easier items but struggle with advanced geometry and combinatorial reasoning. Notably, Gemini 2.0 Flash achieves the highest precision on image based tasks, followed by Qwen VL 2.5 72B and GPT 4o, though none approach human level performance. Fourth, a complementary analysis aimed at distinguishing whether models reason or simply recite reveals that Gemini and GPT 4o stand out for their structured reasoning and consistent accuracy. In contrast, Pixtral and Llama exhibit less consistent reasoning, often defaulting to heuristics or randomness when unable to align their outputs with the given answer options.